AGETS MBR: An Application of Model-Based ... Diagnostics Howard A. Winston and Robert T. Clark’

advertisement
From: IAAI-95 Proceedings. Copyright © 1995, AAAI (www.aaai.org). All rights reserved.
AGETS MBR: An Application of Model-Based Reasoning to Gas lbbine
Diagnostics
Howard A. Winston and Robert T. Clark’
United Technologies Research Center
411 Silver Lane
East Hartford, Connecticut 06108, USA
haw@utrc.utc.com
GeneBuchina
United Technologies - Pratt & Whitney Government Engine and Space Propulsion
I?O. Box 109600
West Palm Beach, Florida 33410-9600, USA
buchina@pwfl.com
Abstract
Currently, a common difficulty in diagnosing failures within
Pratt & Whitney’s FlOO-PW-100/200
gas turbine engine
occurs when a fault in one part of a system, comprising an
engine, airframe, test cell, and Automated Ground Engine
Test Set (AGETS) equipment, is manifested as an out-ofbounds parameter elsewhere in that system. In such cases, the
normal procedure is to run AGETS self-diagnostics on the
abnormal parameter. However, because the self-diagnostics
only test the specifed local parameter, it will pass, leaving
only the operators’ experience and traditional fault isolation
manuals to locate the source of the problem in another part of
the system. This paper describes a diagnostic tool (Le.,
AGETS MBR) designed to overcome this problem by
isolating failures using an overall system troubleshooting
approach. AGETS MBR was developed jointly by personnel
at Pratt & Whitney (PW) and United Technologies Research
Center (UTRC) using an Artificial Intelligence tool called
Qualitative Reasoning System (QRS).
Task Description
Task Performed and Problem Solved
Currently, a common difficulty in diagnosing failures within
Pratt & Whitney’s (PW) FlOO-PW-lOO/200 gas turbine
engine occurs when a fault in one part of a system, comprising an engine, airframe, test cell, and Automated Ground
Engine Test Set (AGETS) equipment, is manifested as an
out-of-bounds parameter elsewhere in that system. In such
cases, the normal procedure is to run AGETS
self-diagnostics on the abnormal parameter. However,
1. Currently with Sun Microsystems-SunService
175 Capital Boulevard
Rocky Hill, Connecticut 06067, USA
Rob.Clark@East.Sun.COM
because the self-diagnostics only test the parameter
specified, it will pass because parameter tests are local tests
that cannot uncover malfunctions in other parts of the
system. At this point, only the operators’ experience and
traditional fault isolation manuals can be used to try to
resolve the problem.
AGETS is used by the U.S. Air Force and
Guard to test, trim and diagnose problems
Whitney FlOO-PW-100 and FlOO-PW-200 jet engines.
Sixty-six AGETS units exist worldwide, and have been in
operation since 1985. AGETS measures over 240
parameters from the engine, connected test equipment, and
itself. These parameters include pressures, temperatures,
rotational speeds, voltages, resistances, and discrete signals.
Many of the monitored parameters originate from engine
sensors, pass through AGETS for measurement, and
continue on to the Engine Electronic Control (EEC) or
Unified Fuel Control (UFC). The presence of the EEC and
UFC complicate the diagnostic process because they can
attempt to compensate for abnormal parameter deviations
and thus mask fault symptoms. In addition, erroneous
signals due to operator error or hardware malfunction can
cause the control system to react to nonexistant problems.
The number of parameters and the potential for control
system interference provide ample opportunity for difficult
problems to arise.
The fault trees found in the FlOO and AGETS paper
maintenance manuals (Intermediate Troubleshooting .. .
1984; Organizational and Intermediate ..* 1988) are ical
of the most widely used approach for aiding maintainers
during troubleshooting. Fault trees present the maintainer
Division
Winston
181
with a sequence of troubleshooting steps designed to isolate
a predefined set of failures. However, the fault tree
approach suffers from a number of shortcomings. In
particular, paper fault trees are:
0 Unable to diagnose faults with novel or unforeseen
symptoms,
0 Inflexible in the sequence of tests required to isolate a
fault,
@ Cumbersome to use, and
* Costly to update.
An alternative (not currently used by AGETS operators)
to the paper-based fault tree is the rule-based expert system.
Rule-based diagnostic expert systems use experts’
knowledge about the relations between symptoms and
causes, encoded in the form of IF/THEN rules. Sequences of
troubleshooting steps are computed dynamically by
performing logical inferences over these rules.
However, as problem complexity increases, development
and maintenance of rule bases can become extremely
difficult and costly. Moreover, since the rules in the
knowledge base are derived from the past experience of
experts, rule-based systems, like fault trees, are limited to
diagnosing faults that have been experienced before and
have well established characteristics.
Pratt 8z Whitney and the Air Force) who could use their
experience to solve global diagnostic problems. In addition,
the resulting large manpower costs were aggrevated by a
high turnover rate that diluted the available experience base.
To address this problem, test procedures in the form of
manual technical orders (‘IDS) were proposed. However, it
was realized that the complexity of subsystem interactions
would make it difficult to assure consistency and
completeness in a set of written TOs for global diagnostics.
Moreover, a flexible procedure was needed to handle different system configurations.
Thus, in order to reduce the costs to the Air Force of
providing telephone support for AGETS global system
problems and as a way to formalize and preserve the
experience base derived from handling these problems, an
automated model-based approach to diagnostics was
desirable. The existence of QRS technology at UTRC and
the intuitive match between its representations and the way
in which engineers conceptualized the structure and
operation of AGETS and other components provided the
motivation to develop AGETS MBR.
Objectives of the Application and Motivation for
an AI Solution
AGETS MBR was developed to assist field engineers and
technicians at Kelly Air Force Base (AFB) in Texas to troubleshoot problems experienced by AGETS users while
testing Fl00 engines. The major part of AGETS MBR is a set
of qualitative models of the testing environment that includes AGETS, the FlOO engine, and various pieces of test
cell equipment. The purpose of AGETS MBR is to
troubleshoot problems encountered during steady-state
engine testing using AGETS. Problems meeting the
following criteria were specifically excluded from AGETS
MBR requirements:
Engine fuel system problems
Engine ignition system problems
Back-Up Control problems
Engine start problems
Engine static problems (i.e., problems in which the
engine is not running because, for example, it cannot
be started)
Engine oil system problems
Subtle engine performance problems (e.g., fuel consumption degradations due to component wear)
Problems for which existing diagnostic procedures
are capable of directly isolating causes, given their
associated symptoms
Technical order or AGETS hardware/software
design problems whose solutions have been implemented
Past problems for which insufficient documentation
exists to positively identify causes
Application Description
AGETS MBR
To overcome the shortcomings of fault-based approaches to
diagnostics, a new approach to diagnostics based on representing normal behavior was used (i.e., qualitative
model-based reasoning).
AGETS MBR is such a
model-based reasoning system that uses qualitative models
of the normal behavior of an engine, airframe, AGETS, and
various pieces of test equipment to troubleshoot problems
by isolating failures to one of these system components. The
qualitative models employed by AGETS MBR were
developed using a patented software system, Qualitative
Reasoning System (QRS), developed at United
Technologies Research Center (UTRC).
Other Solutions
Prior to the development of AGETS MBR, diagnostic
procedures had been developed separately and at different
times for the engine, AGETS, test cell, airframe, and
controller subsystems, but no procedure existed that could
diagnose problems in the overall global system. In effect,
each of the subsystem diagnostics were designed as if that
subsystem existed independently of the others. Thus, there
was no formalized or mechanized process to initially
determine which subsystem(s) could be malfunctioning and
to account for how subsystem interactions shaped fault
manifestations.
The gap caused by the lack of a tie-in between
independent subsytem diagnostics, made it necessary to
train and support a large number of engineers (both from
182
IAAI-95
a Air Force supply system problems (i.e., part procurement-related difficulties)
l
Problems for which AGETS self-diagnostics (or
other troubleshooting procedures in technical orders) are capable of isolating causes
0 Software problems
e Detailed component-level engine troubleshooting
The goal of diagnosis with AGETS Ml3R is to isolate
failures to the FlOO engine, AGETS, airframe, or test cell
subsystems. Although AGETS MRR can troubleshoot
within these modules (their associated qualitative models
are hierarchically constructed from detailed models of their
constituents), existing technical manuals are used for component-level troubleshooting to avoid potential conflicts
with the order of established test procedures.
Three model configurations were developed: (i) uninstalled (encompassing AGETS, FlOO engine, and
M-37/T-20
test cell components), (ii) installed
(encompassing AGETS, engine, test cell, and F- 15 or F-16
airframe components), and (iii) standalone (encompassing
AGETS, engine, and airframe components). The goal of
troubleshooting is to isolate a failure to a major subsystem
(i.e., AGETS, engine, test cell, or airframe) since no
diagnostic aids exist for the global system, except for the
knowledge of experienced troubleshooters. Furthermore,
many problems encountered can seemingly be attributed to
engine or test cell components, but in fact originate in
AGETS. As noted above, more detailed troubleshooting
relies on existing fault isolation techniques due to a desire to
minimize development cost and eliminate confusion in
cases where AGETS MBR might conflict with established
troubleshooting procedures.
functions such as state generation, fault detection, diagnosis,
troubleshooting, and fault tree generation. AGETS MBR, as
delivered to the U.S. Air Force San Antonio Air Logistics
Center at Kelly AFB, consists of the qualitative reasoner
performing the troubleshooting function using qualitative
models of the FlO&PW-100/2OO engine, associated F-15
or F-16 airframe components, test cell equipment, and
AGETS.
To perform troubleshooting, QRS fist uses constraint
propagation on qualitative models to determine whether
given symptoms correspond to a failure. Next, QRS uses
hierarchical constraint suspension (Davis 1984) to
determine which failures could have caused the current
symptoms. For each member of this list of suspected
component failures, QRS generates the predicted values for
model parameters that have not yet been measured or
observed.
In order to perform efficient troubleshooting, QRS uses a
process called Intelligent Test Selection to choose the next
test or observation to request. Intelligent Test Selection uses
knowledge of component failure rates, along with model
predictions, to estimate probabilities of test outcomes and
the extent to which each available test can be expected to
isolate a failure. QRS chooses the test that has the greatest
overall utility, considering such factors as the extent to
which each test is expected to isolate a fault, the a priori
probabilities of various component failures, and the cost of
performing each test.
Qualitative Reasoning System (QRS)
1
This section describes the artificial intelligence tool, QRS,
used to develop AGETS MBR. QRS (Clark, Gallo, & Hamilton 1994) is a software system, developed at UTRC,
designed to support the development of diagnostic
applications of qualitative reasoning. Qualitative reasoning,
or more generally symbolic model-based reasoning, is a
subfield of artificial intelligence concerned with the
computation of possible behaviors of a device from a
qualitative model of its structure and function. A qualitative
model is an abstract representation of a device that allows
decisions to be made from a high-level understanding of a
situation, without the need for specific quantitative details
that may be either (i) unavailable, (ii) misleading (because
the device might be broken so that the precision of a
quantitative model might be inappropriate), or (iii) untimely
to attain.
QRS is made up of two major components: the qualitative
model developer and the qualitative reasoner. The model
developer is a graphical user interface which helps
application domain experts build and test qualitative models. The qualitative reasoner is responsible for determining
behaviors of qualitative models and can perform many
Gener-
Phase
Gener-
ator
ator
cz.
A Statns
Cl
I
G2
I
On
1 Off
1
Gl
1
1 Off
1
1
Gl
1 Off
1 Off
1 On
1 62
1 On
On
trO1
Phase
B Status
Phase
C Stat&Is
On
1 Off
1 On
1 Off
1 On
1
Table 1. Normal value assignments for CIRCUIT-I.
A SimpleExample.In this section, we describe a simple example to illustrate the troubleshooting process in QRS. The
example is based on a portion of a generic electrical power
distribution subsystem. The example system, designated
CIRCUIT-l, is depicted in Figure 1. There are three inputs
and three outputs to the system, as follows:
0 GI, generator # 1 power status (On, Off}
e G2, generator # 2 power status (On, Off}
e K2 Control, controls whether the K2 contactor selects Gl or G2 (Gl, G2)
e Phase A Status, shows the status at Phase A of the
load (On, Off}
Winston
183
Power Source
Bus Between
, K4andK2
,
No. 2 Junction
Gl
Phase A
Phase B
Phase C
An
Box
U,“,,,,
rI
G2
Phase A
Figure 1. Schematic for CIRCUIT-l,
0 Phase I3 Startus, shows the status at Phase B of the
load (On, Off}
0 P&e C Status, shows the status at Phase C of the
load {On, Off)
A description of the normal behavior of the system is as
follows: If GI is On and K2 Control selects GI, then all
phases should be On. If G2 is On and K2 Control selects G2,
then all phases should be On. Otherwise, all phases should
be Ofi From the normal behavior models of CIRCUIT-l,
the QRS state generation algorithm can generate the possible normal behaviors of the system. These normal states
are summarized in Table 1.
Diagnosis. Suppose that an operational checkout procedure
is being performed on CIRCUIT-l. The first step in the
checkout is to power on Generator Gl, power off Generator
G2, set K2 Control to select Gl, and observe the Phase A, B,
and C Status Indicators. As seen in the third row of Table 1,
normal indications for this condition are Phases A, B, and C
On. Suppose, however, that Phases A and B are On and
Phase C is Off. Using constraint propagation, QRS can determine that these symptoms are inconsistent with the normal behavior of the system.
QRS next uses hierarchical constraint suspension to determine which component failures can account for the symptoms. Assuming a single point of failure, QRS identifies the
following
components
as suspects: Load, Wire
H6381A10C, Current Limiter CL9, Wire H6448AKX, K2
Contactor, Phase C Bus, or Generator G 1.
To better understand how QRS determines these suspects,
consider the two components Current Limiter CL9 (a suspect) and Current Limiter CL7 (not a suspect). To test Current Limiter CL9, its constraints are temporarily removed
(suspended) from the constraint network. QRS then determines if a legal state can be generated with this new
constraint network and the original failure symptoms. Sus184
IAAI-95
an electrical power distribution subsystem.
pending the constraints of CL9 removes the constraint between the input and output values (i.e., voltages and currents) of CL9. Thus, the symptom that the Phase C Status
Indicator is Off is consistent with the constraint networkcorresponding to a CL9 failure, and Current Limiter CL9 is
identified as a suspect.
The constraint suspension process is also used to test the
Current Limiter CL7 failure hypothesis. After suspending
the constraints associated with CL7, QRS attempts to generate a legal state consistent with the symptoms. This time,
however, the conflict between the symptoms and the
constraint network remains (i.e., a CL7 failure cannot account for Phase C status being Off). Since a legal state cannot be generated for the CL7 failure hypothesis, CL7 is not
identified as a suspect.
Fault Isolation. As mentioned previously, to isolate the failure to a single component (or minimal set of components),
QRS uses a process called Intelligent Test Selection that enables QRS to compute the utility of each applicable test. A
test’s utility is defined as its diagnostic power (i.e., degree of
fault isolation) divided by its associated cost.
For example, assume there are two tests that may be performed on CIRCUIT-l. TEST-1 measures the voltage at
terminal C3 of the K2 Contactor. TEST-2 measures the voltage between Wire H6381AlClC and the Load. Given the initial list of suspects, TEST-l partitions the list as follows: If
the voltage is normal, then Generator G 1, Phase C Bus, and
K2 Contactor are eliminated; If the voltage is zero, then Wire
H6448AlOC, Current Limiter CL9, Wire H6381AltX and
Load are eliminated. TEST-2 partitions the list of suspects
as follows: If the voltage is normal, all components except
the Load are eliminated; If the voltage is zero, the Load is
eliminated.
To determine which test to select, the utility of each test
must be computed. Assuming equal test costs and equal a
priori component failure probabilities, TEST-l would have
a higher utility than TEST-2. Intuitively, this is because
TEST-l partitions the remaining hypotheses into two setsof
approximately equal size. Therefore, TEST-l will remove
about half of the uncertainty of the diagnosis, regardless of
the outcome of the measurement. In contrast, since TEST-2
partitions the remaining hypotheses into two sets of unequal
size, TEST-2 is most likely to remove just one seventh of the
uncertainty of the diagnosis.
In practice, the costs of different tests and the failure probabilities of various components may vary widely. Thus, the
computed utility of TEST-2 may actually be higher than that
of TEST-l. For example, if we assume that TEST-2 is very
inexpensive compared to TEST-l, then TEST-2 may be selected before TEST-l, even though TEST-l has greater
diagnostic power. Alternatively, if the failure probability of
the Load is large relative to the failure probabilities of the
other six suspects, the diagnostic power of TEST-2 may be
greater than that of TEST-l. The process of computing the
test utilities and selecting the most cost-effective test is repeated until only one hypothesis (or minimal set of hypotheses) remains.
Modeling Methodology
AGETS, FlO engine, airframe, and test equipment were
modelled with qualitative variables, value spaces, and confluences. Qualitative variables differ from quantitative
variables in that numerical values are not required. Instead
qualitative variables normally have values drawn from value
spaces such as {Positive, Zero, Negative) or {High, Low,
Normal}. For properties requiring a greater level of detail,
additional landmarks can be inserted in the value space between Zero and Infinity. Landmarks can be used to represent
numerical limits, values that bound regions of qualitatively
distinct behavior, or simply on and off parameter settings.
Qualitative variables are then combined into qualitative
equations, or confluences, to describe normal device behavior. Confluences look very much like normal numerical or
algebraic equations, except that parameters of unlike materials (e.g,. fuel and air) and unlike properties (e.g., pressure
and temperature) can be directly related with qualitative operators. (See (de Kleer & Brown 1984; de Kleer & Williams
1987; de Kleer 1993; Forbus 1984; Forbus 1993; Kuipers
1984; Kuipers 1986; Kuipers 1993) for more information
about qualitative modeling, representation, and reasoning.)
The objective of modeling AGETS and other system
components was to construct a functional representation, not
necessarily a physical one. That is, logically related components were sometimes grouped together, even though they
might exist in physically different parts of the overall system. For this purpose, QRS supports the representation of
primitive components (without substructure) in the form of
elementary qualitative models and complex corn ents
(having an internal structure) in the form of compound qualitative models. Functional groupings of related components
were represented as compound models in AGETS MBR.
Three steps occur in the elementary model building
process: (i) identify relevant parameters, (ii) determine
which parameters are also input and output terminals, and
then (iii) constrain the parameters with confluences. As an
example, for a burner model, the relevant parameters are Air
Flow In, Air Pressure In, Air Temperature In, Fuel Flow In,
Gas Flow Out, Gas Pressure Out, and Gas Temperature Out.
All of these parameters are terminals. The confluences are
as shown in Figure 2(a).
Compound model construction simply involves
connecting terminals of like material and property between
elementary or compound models, as shown in Figure 2(b).
(Forbrevity,AirInorOutisanabbreviationforthethreeAir
Flow, Temperature, and Pressure In or Out parameters.) This
connection of input to output terminals continues with
elementary and compound models until the toplevel application model is completed. Unlike the ability to directly
relate or equate parameters of dissimiliar materials in
confluences, only terminals comprised of the same material
can be connected because they essentially establish identity
relations between parameters in different models.
LessonsLearned
From our experience designing and building AGETS MBR,
several insights were gained about the application of AI
technology.
1) The acceptance of AGETS MBR was
an important step taken at the start of
In order to insure that the features and operation of
the delivered system would meet the customers’
expectations, we conducted a two day JAD (Wood
$ Silver 1989) (joint application design) session at
Kelly AFB in order to understand what the Air
Force needed, how they wanted it delivered, and to
Figure 2. Elementary and Compound Models.
Winston
185
precisely define the application’s capabilities. JAD
is a brainstorming process for eliciting a set of
project requirements and specifications from a
customer with only minimal vendor interference.
(UTRC technologists were only permitted to
answer technical feasibility questions.) This
helped to preclude proposing solutions that
modified the “nail to fit the hammer”.
2) The project was divided into two phases that
minimized risk to the customer. The first phase
developed a prototype system for a small AGETS
subset that was demonstrated to the customer and
checked against requirements that arose from the
initial JAD session. New requirements also arose
from interaction with the prototype. This insured
that the complete AGETS MBR diagnostic system
delivered at the end of phase two met or exceeded
Air Force expectations.
3) AGETS MBR was developed on a SUN platform
and delivered on a Pentium-based PC. Although
the customer assigned no resource limitations to
therequiredPCdelivery platform, theperformance
of the delivered system was, nevertheless, slower
than the Air Force expected and the system
probably lost some acceptance due to this fact. We
have learned that the performance of an AI
application is one of its most important features.
Application Use and Payoff
Results
AGETS MBR has been in use by members of the San Antonio Air Logistics Center and employees of Pratt & Whitney
since June of 1994. We tested AGETS MBR on the set of all
telephone hotline calls received at Kelly AFB from worldwide AGETS field locations from 6/l/94 to 10/12/94. There
were a total of 31 calls received, Of these calls, 13 were
acceptable for AGETS MBR troubleshooting as outlined in
the Section labelled AGETS MBR. In all 13 of these cases,
given the initial symptoms, AGETS MBR detected
discrepancies between the expected behavior and the actual
behavior of the system under test. Additionally, each initial
fault hypothesis list generated by AGETS MBR included the
actual cause of the problem.
Due to the conditions outlined in the AGETS MBR Section, the installed version of AGETS MBR is not used to
isolate the failure beyond the highest level modules in the
system under test. To understand the true power and
capabilities of the qualitative models and qualitative
reasoner employed by AGETS MBR and to eva.lua& the
effectiveness of AGETS MBR (i.e., what the system could
do if not for the limited scope of how it is routinely used), the
13 cases identified above were put through a more rigorous
analysis. For these tests, AGETS MBR was permitted to
diagnose down to the Line Replaceable Unit (LRU) level of
the system under test. The results of these tests showed
whether AGETS MBR could localize the actual failure to
one of its elementary models, as well as the final level of
ambiguity reduction from the fault hypothesis list generated
from the given initial symptoms.
Table 2 summarizes the results of these detailed tests.
Diagnostic coverage is defined as whether or not the actual
cause was listed as a possible fault identified by AGETS
MBR. In other words, the diagnostic coverage is positive if
AGETS MBR could identify the actual failure, and negative
otherwise. AGETS MBR was able to do this in 12 of the 13
cases,leading to an overall diagnostic coverage of 92%. The
other important statistic to be seen from examining the table
is that the average ambiguity reduction is 80%. Ambiguity
reduction is defined as the percentage of components that are
removed from consideration as possibly causing the fault
given the symptoms as defined in the case. These statistics
are displayed graphically in Figure 3.
As of 4/15195, a total of 77 telephone calls were received,
of which 23 were within the design scope of AGETS MBR.
Faults were detected in all 23 cases, and the preliminary hypothesis list included the actual cause in all these cases.
Payoff
Because of the low frequency of trouble calls and the decision to limit diagnosis capability to high-level system modules, the benefits of AGETS MBR are hard to quantify. As
such, efforts are underway to relax the operational require-
Remaining
Possible Hypotheses 20%
Incorrect Hypotheses 8%
Coverage
92%
Reduction
80%
Figure 3. Diagnostic coverage and ambiguity reduction.
186
IAAI-95
ments of AGETS MBR in order to diagnose to the LRU level. Post-delivery experience has shown that many difficult
problems occur within AGETS. Allowing AGETS MBR to
troubleshoot such problems would make payoff calculations
in terms of financial and/or labor metrics more tractable.
An unexpected benefit of system delivery was the decision to base the PW development engineer at Kelly AFB as
an on-site representative. The experience of working with
QRS and AGETS MBR enhanced his troubleshooting skills
by forcing him to look at familiar situations in unfamiliar
ways. Many times during development, experts learned
something new about troubleshooting AGETS andFlO0 engines. PW Government Engine and Space Propulsion (PW/
GESP) also accomplished the goal of developing experience
in artificial intelligence diagnostic systems. Specifically, an
engineer acquired expertise in modeling complex systems
for making accurate diagnoses. PW/GESP plans to utilize
this expertise in future projects.
Application Developmentand Deployment
Development Processes
The basic model development process iterated testing and
redesign processes between UTRC and PW. Figure 4 shows
how model development steps were coordinated between
these two organizations. PW provided initial model designs
based on their domain knowledge which were reviewed by
UTRC in order to ensure consistency with other models and
compatibility with QRS’s modeling primitives. PW then
coded the models into QRS and tested their fault detection
capability (i.e., shallow testing). Fault detection only
determined whether or not faults existed based on canonical
sets of corresponding symptoms.
If a review by PW of shallow testing results was
satisfactory, UTRC proceeded with detailed hypothesis
testing (i.e., deep testing). Hypothesis testing determined
whether diagnostic procedures applied to the models could
find correct hypothesis sets. Based on PW’s review of deep
testing results, further model debugging was done at UTRC.
PW then proceeded to redesign and recode the new models.
If a review by PW of shallow testing results was
unsatisfactory, PW would debug, redesign, and recode the
models. This procedure iterated until PW determined
the results of deep testing were satisfactory.
The AGETS MBR models were developed over the
course of approximately 18 months and at a cost of 15 manmonths involving two developers at any given time (one
fiorn UTRC and one from PW). AGETS MBR encompasses
three AGETS configurations (unique models are specific to
a particular configuration):
0 Installed, comprising 1398 elementary and 242
compound models. 212 models are unique.
0 Uninstalled, comprising 1368 elementary and 229
compound models. 18 1 models are unique.
@ Standalone, comprising 1362 elementary and 235
compound models. 209 models are unique.
An existing model of a commercial turbofan engine
(Winston et al. 1991) was used as the starting point for the
gas path portion of the FlOO engine model. The confluences
were derived from the basic thermodynamic behavior of gas
turbine components such as compressors, burners, and
turbines. The electrical and hydromechanical control of the
FlOO proved to be especially challenging to model. Many
iterations were required to eliminate static instability in the
Figure 4. UTRC/PW Software Development Process.
Winston
187
control system model. The AGETS models were designed
from electrical schematics in the AGETS maintenance
manual (Organizational and Intermediate . .. 1988). While
AGETS contains numerous components from wires to
computer cards, similarity permitted a high degree of model
reuse. Additionally, software routines were used to
automate construction of large compound models. A small
number of aircraft and test cell component models were
developed at relatively small expense.
After a model was constructed, it was tested for accuracy
using the State Generation and Diagnosis utilities in QRS.
First, the number of legal states generated for a set of input
conditions was determined. In most cases, exactly one
generated state was desirable. In certain cases, more than
one legal state was needed to accurately describe the
function of a component. After successful completion of
State Generation testing, typical system problems were
simulated, first to ensure fault detection (i.e., shallow testing), then to compile hypothesis lists (i.e., deep testing).
These hypothesis lists were checked against historical
evidence and reviewed by domain experts. As explained
above, deviations from expected results usually caused an
iteration in the model design.
Test procedure design and implementation was
accomplished concurrently with model testing. The
majority of test procedures used by AGETS MBR were of
two main types: parameter observations and AGETS
hardware checks. A small number of test procedures
involved engine observations. As before, a desire to avoid
user confusion by conflicting
with established
troubleshooting methods was the reason only a few
engine-related test procedures were utilized.
The final process in the AGETS MBR development was
converting the entire system from a UNIX platform, upon
which it was developed, to an IBM+zompatible PC platform.
AGETS MBR was delivered on an IBM-compatible PC
utilizing a 60 MHz Pentium processor and 64 Mb of RAM.
The PC platform was chosen at the customer’s request for
commonality with existing hardware.
Deployment Process
The AGETS MBR system was intended to be easily used by
personnel at the San Antonio Air Logistics Center
(SA-ALC) at Kelly AFB. It was designed to be user-friendly by incorporating an easy to use graphical user interface
along with a full-featured on-line help system. Because of
these objectives, the deployment process was similar to that
of a commercial software package.
AGETS MBR was successfully installed at the SA-ALC
in June of 1994. Installation consisted of procuring the
hardware, obtaining needed software licenses, and loading
the completed QRS models. Following installation, a two
week acceptance test and training period was completed to
the customer’s satisfaction. Sample and actual problems, as
well as customer test cases, were used to qualify the system
during the acceptance test. Training consisted of supplying
user’s manuals and conducting tutorial sessions. After
188
IAAI-95
completion of training, an employee of Pratt $ Whitney who
could answer questions about AGETS MBR remained onsite at SA-ALC.
Table 2. Detailed Results.
Maintenance
As of this writing, there have been two updates to the
AGETS MBR software. The first update involved miscellaneous user-interface improvements while the second update
involved a major performance improvement. Neither of
these updates were planned as maintenance, and new releases were not included in the customer’s original contract.
As such, new updates are not planned at this time. However,
several improvements to the AGETS MBR software have
been identified by the customer. These enhancements include extending coverage to previously excluded systems
(e.g., engine fuel system, oil system, etc.), inclusion of decision trees automatically compiled from AGETS MBR models for more efficient performance, as well as an extended tutorial system, and field deployment of the AGETS MBR
!SOftWiUtL
Pratt & Whitney complex equipment support engineers
are responsible for maintaining AGETS MBR qualitative
models - the knowledge base that supports diagnosis of the
actual system. Because qualitative models are based on normal device behavior which does not often change over time,
it is not expected that many changes will be required to the
knowledge base. This is in contrast to fault-based approaches (e.g., fault trees or shallow rule-based systems)
where new failure modes can be frequently discovered.
However, when changes to the knowledge base are needed,
due to incorrect models of devicebehavior or changes to the
actual system, updating the application should be fairly
straightforward in that one qualitative model can be easily
substituted for another qualitative model without disrupting
the remaining models in the system.
de Kleer, J.; and Brown, J.S. December 1984. A
tative
Physics Based on Confluences. Artijkial Intelligence
X17-83.
Conclusion
de Kleer, J.; and Williams, B.C. April 1987. Diagnosing
Multiple Faults. Artificial Intelligence 32( 1):97-130.
In many cases, problems experienced during FlOO engine
testing with AGETS are very difficult to troubleshoot due to
any or all of the following:
e the volume and complexity
of AGETS
measurements,
0 possible interference by the engine’s electrical and/or
hydromechanical control systems, and
* the lack of a formal troubleshooting aid for the testing
system.
AGETS MBR, utilizing QRS software, provides support
personnel at Kelly AFB with a diagnostic aid which
encompasses the entire engine/airliame and testing system.
In addition, AGETS MBR does not suffer from the many
pitfalls of traditional fault trees. These considerations and
the results of AGETS MBR testing motivate the shift from
fault tree-based approaches to normal behavior approaches,
as employed in AGETS MBR.
Acknowledgements
The authors would like to thank the following people who
contributed to the successof the AGETS MBR system. Tom
Hamilton has led UTRC’s QRS technology development for
over 10 years and helped to establish and define the AGETS
MBR program. At Pratt & Whitney, Eric Meyer was a lead
engineer on the AGETS program who initiated the AGETS
MBR collaboration with UTRC and contributed to its early
development, Scott Connally provided valuable information
about AGETS operation, and Bob Wilkoff developed a
complementary diagnostic case-based system for handling
variations of known failure modes. Don Wade (Pratt &
Whitney) moderated the early JAD session at Kelly AFB.
References
Clark, R.T.; Gallo, S.; and Hamilton, T.I? January 1994.
Qualitative Reasoning: Theory and Applications. In Proceedings of the 32nd AIAA Aerospace Sciences Meeting.
Reno, Nevada.
Davis, R. December 1984. Diagnostic Reasoning Based on
Structure and Behavior. ArtijZial Intelligence 24:347-410.
de Kleer, J. February 1993. A View on Qualitative Physics.
Art@cial Intelligence 59: 105-114.
Forbus, K. December 1984. Qualitative Process Theory. Artificial Intelligence 24~8%168.
Forbus, K. February 1993. Qualitative Process Theory:
Twelve Years After. Artificial Intelligence 59: 115-123.
Hamilton, TX July 1988. HELIX: An Application of
Qualitative Physics to Diagnostics in Advanced
Helicopters. International Journal for AI in Engineering
3(3): 141-150.
Kuipers, B.J. December 1984. Commonsense Reasoning
about Causality: Deriving Behavior from Structure. Art&%
cial Intelligence 24: 169-203.
Kuipers, B.J. September 1986. Qualitative Simulation. Artificial Intelligence 29(3):289-338.
Kuipers, B.J. February 1993. Qualitative Simulation: Then
and Now. Artificial Intelligence 59: 133-140.
Intermediate Troubleshooting and Engine Subsystem
Engine.
USA.F
Model
Description,
A-imaft
FlOO-PW-100(3). July 1984. Technical Order, T.G.
2J-FlOO-21-3, WI’ 052.
Organizational and Intermediate
tenance Automated
Ground Engine Test Set PWA52105. August 1988. Technical Order, T.O. 33D4-6-690-2.
Winston, H.; Sirag, D.; Hamilton, T.; Smith, H.; Simmons,
D.; and Ma, I? January 1991. Integrating Numeric and
Symbolic Processing for Gaspath Maintenance. In Proceedings of the 29th AIAA Aerospace Sciences Meeting. Reno,
Nevada.
Wood, J., and Silver, D. 1989. Joint Application Design.
New York, NY: John Wiley 8z Sons.
Winston
189.
Download